Standard NLP models are trained on majority-White internet text. They were never designed to read how BIPOC, LGBTQ+, and first-generation youth communicate distress — and they consistently fail to. CulturalBERT-VLAP was built specifically for those communities. Not adapted. Built.
On high-distress signal detection in active IRB validation with the University of Maryland. Results to be published upon study completion.
General-purpose NLP models — including large language models — are trained on internet text that skews heavily toward majority-White, educated, English-speaking populations. The result is a systematic blind spot: the specific ways BIPOC, LGBTQ+, and first-generation youth signal emotional distress are consistently misread, deprioritized, or missed entirely.
This isn't a model failure in the traditional sense. These models perform well on the populations they were trained for. The failure is in deploying them, uncorrected, for populations they were never trained on — and expecting accurate clinical signal detection to follow.
The gap isn't theoretical. In clinical practice, it means a youth saying "lowkey been struggling fr" is read as casual. A pre-disclosure minimization pattern ("it's not that deep but...") reduces the model's confidence instead of increasing it. Community-developed coded language — terms built specifically to avoid content filters — registers as noise. Standard models consistently miss these signals in the communities that most need them caught.
CulturalBERT-VLAP is a BERT-architecture language model fine-tuned on a purpose-built corpus of culturally specific mental health language. The base BERT architecture was selected for its bidirectional context processing — essential for reading the layered meaning in code-switching, minimization patterns, and culturally framed expressions where individual words carry different weight depending on surrounding context. The model was then extended with a culturally specific vocabulary and fine-tuned against a clinically annotated training corpus.
The bidirectional encoding layer processes the full context window simultaneously in both directions — unlike unidirectional models that read left-to-right. This is architecturally necessary for cultural signal detection: the meaning of "lowkey" depends entirely on what follows it, and the clinical significance of a minimization hedge ("it's not that deep but") only becomes clear in context of what comes after the conjunction.
Standard BERT vocabulary was extended with 2,400+ AAVE terms, youth vernacular expressions, code-switching patterns, and community-developed coded language — including terms built specifically to circumvent content filters (e.g., "unaliving"). This extension was compiled through community engagement with BIPOC and LGBTQ+ youth populations and validated by licensed clinicians with relevant community competency. Without this extension, approximately 23% of the training corpus would be processed as unknown tokens by the base model.
The extended BERT model was fine-tuned on 198,000+ annotated language samples drawn from the communities VLAP serves — not web-scraped text, but specifically collected and clinically annotated mental health communication. Each sample was labeled by licensed clinicians with community competency training against the 42-signal taxonomy across five behavioral dimensions. The fine-tuning process was conducted in multiple phases with inter-annotator agreement validation.
The model output layer generates dimensional signal profiles across the five-dimension taxonomy — not diagnostic outputs, not severity scores, and not clinical recommendations. The output is an interpretive context package: which signal patterns are present, their dimensional classification, and cultural interpretation notes that help clinicians understand what the patterns typically indicate in the communities that produce them. Every output includes a non-diagnostic disclaimer and is designed to support clinical judgment, not substitute for it.
VLAP processes member language in-memory. Verbatim input text is not stored after signal profile generation. The retained output is a dimensional signal profile — not a transcript, not a quote, not a record of the member's exact language. This architectural constraint is a HIPAA technical safeguard, not a configurable setting. It cannot be disabled by organizational administrators or overridden by clinical staff.
The training data for CulturalBERT-VLAP was not web-scraped, crowd-sourced generically, or assembled from existing open datasets. It was specifically collected from and with the communities the model serves — and annotated by licensed clinicians who have worked within those communities.
Language samples were collected in partnership with community organizations serving BIPOC, LGBTQ+, and first-generation youth populations — not scraped from public platforms. Consent protocols, anonymization procedures, and community review were built into the collection process.
All training samples were annotated by licensed clinicians with documented community competency training in the relevant population cohorts. Inter-annotator agreement was validated across the annotation cohort before samples were included in training data.
The 2,400+ vocabulary extension was compiled through structured engagement with youth from the target communities — not through generic web scraping. Vocabulary candidates were reviewed for clinical accuracy by the annotation cohort before inclusion.
False positive and false negative rates were disaggregated by demographic subgroup throughout training — not as a post-hoc audit but as an operational gate. Signal detection performance must meet minimum parity thresholds across BIPOC, LGBTQ+, and first-generation subgroups for a model version to be deployed.
The VLAP signal taxonomy was developed through clinical annotation of the training corpus — identifying recurring patterns of culturally framed distress expression that standard clinical instruments and general NLP models systematically miss. The taxonomy is not static. It is updated as language evolves, with community input informing additions and revisions through a structured clinical review process.
Indirect hopelessness, temporal compression, and "no point" framing expressed through AAVE, coded youth language, and minimization patterns. Standard models read these as low-confidence signals. CulturalBERT-VLAP reads them as the primary signal. The cultural norm of underreporting — particularly prevalent in BIPOC male youth populations — means direct expressions of hopelessness are less common than coded expressions. HOP signals are most clinically significant when combined with CCM modifiers.
Withdrawal signals, relational distancing, and absence of connection framed as normal or preferred. ISO signals are particularly significant for BIPOC and LGBTQ+ youth whose support networks are often culturally specific and difficult to replace — the stakes of isolation are higher. Critically, ISO signals frequently appear as positively framed statements ("I'm good, just been staying to myself") that read as fine to general models but register as relational withdrawal in cultural context.
Coded self-harm and suicidal ideation language — including community-developed terms built specifically to circumvent content filters. The SHA dimension required the most extensive vocabulary extension: terms like "unaliving," "sewerslide," and others circulating within youth online communities to communicate suicidal ideation while avoiding automated detection. Standard models with standard vocabularies have zero coverage of these terms. SHA-category signals are the highest-stakes VLAP detections and receive the most conservative threshold settings — lower false negative rates, higher false positive tolerance.
Acute distress patterns including perceived burdensomeness, escalating hopelessness with sincerity markers, and suicidal ideation confirmed by authenticity escalators. CRS signals operate as a multiplier on co-occurring signals — a SHA signal becomes a clinical priority when paired with a CRS sincerity marker. CRS-category detections trigger immediate surfacing to Vasl's licensed clinical supervisor team with a 90-minute human response SLA. No automated action is taken; human clinical judgment determines every response.
The CCM dimension is the most architecturally significant of the five — and the most clinically important. CCM signals don't indicate distress directly; they modify how other signals should be interpreted. Pre-disclosure minimization (CCM-04) paired with an ISO signal means the withdrawal is more significant than it appears. Spiritual deflection (CCM-12) may indicate either genuine resilience or disclosure avoidance. Code-switching (CCM-08) indicates the member is speaking to a perceived authority audience — context that changes what they're willing to disclose. Standard models have no equivalent to the CCM dimension; they read utterances at face value without cultural framing.
VLAP's accuracy claims are grounded in active IRB-approved clinical research with the University of Maryland — not internal testing, not synthetic benchmarks, and not general NLP performance metrics that don't account for cultural signal specificity.
VLAP's ability to detect high-distress signals when they are present in member language — measured against clinician-adjudicated ground truth in the IRB study cohort. Sensitivity is optimized conservatively for SHA and CRS dimensions: we accept more false positives to minimize missed crisis signals.
Percentage of the VLAP training corpus that would be processed as unknown tokens by a standard BERT model without the extended vocabulary — representing the portion of culturally specific language that standard models are structurally unable to read.
VLAP signal accuracy is being validated through an IRB-approved clinical study with the University of Maryland, using production deployment data from live Vasl cohorts. The study compares VLAP signal output against clinician-adjudicated gold-standard assessments of the same member language. Results will be published in peer-reviewed literature upon study completion.
False positive and false negative rates are disaggregated across BIPOC, LGBTQ+, and first-generation subgroups in both the training validation and the IRB study. Parity thresholds are enforced operationally — a model version that meets aggregate accuracy targets but fails subgroup parity is not deployed. Bias monitoring is an ongoing production gate, not a one-time evaluation.
Sensitivity measures how consistently VLAP detects signals when they are present — not the rate at which all surfaced signals are clinically significant in a given instance. A high-sensitivity threshold means more signals are surfaced, which is appropriate for a clinical support tool. The clinical significance of any specific signal is always determined through human clinical review, not by the model.
The active IRB study is currently in the data collection and preliminary analysis phase. Results will be published in a peer-reviewed journal upon completion. The study protocol and preliminary design documentation are available to institutional evaluators under NDA. Contact clinical@vaslhealth.com to request access.
Vasl Health's clinical validation approach is overseen by its Senior Medical Advisor, Panagis Galiatsatos, MD, MHS — Assistant Professor of Medicine at Johns Hopkins University School of Medicine. Dr. Galiatsatos provides clinical oversight on VLAP's signal detection methodology, accuracy validation approach, and non-diagnostic output framing.
VLAP is a clinical decision-support tool, not a member-facing AI. It operates entirely behind the clinical layer — invisible to the people whose language it processes. Every signal it surfaces is directed to a licensed clinician or certified coach, reviewed by a human, and responded to through human clinical judgment. The platform is built so that automated action in response to a clinical signal is architecturally impossible.
VLAP processes only language shared through Vasl's care channels — daily check-ins and coach messaging threads. Peer group posts, external social media, school email, and any other channel are not processed.
CulturalBERT-VLAP processes the language against the 42-signal taxonomy. In-memory only — verbatim text is not retained after processing. Output: a dimensional signal profile.
Coaches see a simplified surface of VLAP output in the AI Client Insights panel: plain-language pattern alerts and mood trajectory summaries for their active members. No dimensional codes, no clinical jargon.
When a member is connected to a licensed clinician, the pre-session view includes the full VLAP dimensional signal profile — dimensional codes, pattern descriptions, cultural interpretation notes, and coaching context. Accessible only to licensed clinicians.
Crisis Risk Signal detections are surfaced immediately to Vasl's licensed clinical supervisor team. A licensed clinician reviews the signal and determines the appropriate response. No automated action. Human judgment initiates every response.
VLAP processes the most sensitive category of user data — mental health language from youth in underserved communities. The security architecture was designed specifically for HIPAA-regulated, school-based, and community health deployment contexts. Every constraint below is architectural, not configurable.
VLAP processes member language in-memory. Verbatim input text is not stored after signal profile generation. The retained output is a dimensional signal profile — not a transcript, not a quote. This is a HIPAA technical safeguard, not a configurable setting.
HIPAA Security Rule technical safeguards implemented across all platform components — encryption in transit and at rest, access controls, audit logging, and automatic logoff. Business Associate Agreement required for all organizational deployments. Annual third-party security audit.
Annual SOC 2 Type II audit covering security, availability, and confidentiality trust service criteria. Full audit report available to institutional evaluators under NDA. Audit conducted by an independent third-party auditor.
Individual VLAP signal context is accessible only to the assigned coach (AI Client Insights summary) and the assigned licensed clinician (full dimensional profile). School staff, org administrators, and Vasl team members outside clinical supervisory functions have zero access to individual signal data — architecturally enforced.
For school district deployments, Vasl operates as a direct service provider to students. Student health data generated in Vasl is classified as health information under HIPAA — not as an education record under FERPA — and is structurally inaccessible to school administrators under any circumstances.
Population-level aggregate signal trends surfaced to org administrators use minimum cohort size enforcement to prevent de-identification by inference. Individual member contributions to aggregate data are never discernible. This constraint applies to all org-level reporting, without exception.
VLAP's clinical credibility is grounded in active institutional partnerships — not aspirational affiliations or advisory relationships that don't involve actual work. The partnerships listed below involve ongoing operational collaboration, active research, or formal clinical oversight.
Vasl Health is an active participant in the Mayo Clinic Platform_Accelerate program, which provides access to de-identified patient data infrastructure and clinical research partnership. Vasl is engaged in a two-year data collaboration — Year 1 focused on behavioral health signal model development using longitudinal HL7 FHIR-formatted data (BH encounters, ICD-10 codes for mood/anxiety/trauma/SUD, standardized assessments, demographics/SDoH), with Year 2 targeting prospective validation of CulturalBERT-VLAP signal accuracy against clinical ground truth. Primary contact: Asia Smith, MPH, Program Success Manager.
An IRB-approved clinical study is in progress with the University of Maryland validating CulturalBERT-VLAP's signal detection accuracy against clinician-adjudicated ground truth assessments. The study uses production deployment data from live Vasl cohorts. Results will be published in a peer-reviewed journal upon completion. The study represents the first formal independent validation of VLAP's culturally specific signal detection capabilities.
Panagis Galiatsatos, MD, MHS — Assistant Professor of Medicine at Johns Hopkins University School of Medicine — serves as Vasl Health's Senior Medical Advisor. Dr. Galiatsatos provides clinical oversight on VLAP's signal detection methodology, accuracy validation approach, non-diagnostic output framing, and the clinical governance of the platform's care coordination model. His advisory role involves active participation in clinical review, not nominal affiliation.
Vasl Health provides full technical documentation to qualified institutional evaluators — health systems, research institutions, school district technology teams, and health plan medical directors. All documentation is available under NDA. Contact clinical@vaslhealth.com or use the form below to initiate an evaluation request.